Hey, it’s Alex 👋
What a week! Gemini 4 Argon was Google’s best model for exactly two weeks. Then Google shipped Gemini 4 Carbon, and it’s the new number 1 in the world. Google is SO back. (when they actually release, it’s waitlisted so far)
The other big story is Anthropic’s S-1. For the first time, we get real, audited numbers on what a frontier AI lab looks like as a business. Peter Gostev stayed up making 60+ charts, and we walked through them together on the show.
We also had models trying to escape their sandboxes, and Genie 4, which made me think (again) that we live in a simulation.
Wolfram and I hosted from the same room for once (meatspace!). Peter, Nisten and LDJ joined remotely. Mert Gulsun from Artificial Analysis joined to talk about their new index, and for From the Forge we had the CoreWeave engineers who are “forward deployed” inside the Aston Martin F1 team! What does that mean? We dive into FDE
Google is SO back
Gemini 4 Carbon - #1 on Artificial Analysis, 2 weeks after Argon (X, X, Sundar)
Gemini 4 Carbon scores 59 on the Artificial Analysis Intelligence Index. That makes it the top model in the world on their brand new index, 1 point ahead of Claude Opus 5.5. It also debuts at #1 on Text Arena and #2 on Agent Arena.
Two big models back to back is unusual for Google, after a long stretch of Flash models. There’s no version number either. It’s not Gemini 4.1, it’s just... Carbon. I have no idea why 😅
I haven’t seen a Gemini eval chart like this in a long time. Google is finally competing with the full frontier (GPT-6 Astra, Opus 5.5, Fable 5.1), and on most rows it wins.
My favorite eval here is Vending Bench from Andon Labs, where the model runs a real vending machine business. Carbon finished with $16,484 in the bank. Opus 5.5 finished with about 9K. LDJ was most impressed by RiemannBench, a very hard math benchmark, where Carbon beats Opus 5.5 by 10 points. Math was supposed to be OpenAI’s turf!
Google is pitching Carbon for agentic coding and cybersecurity. The launch price is $2 per million input tokens and $10 per million output, which already includes a 50% launch discount. Wolfram loved this. Models are finding security holes that have been there for years, and fixing them takes a LOT of tokens. In his words: “the cheaper the model that can do this, the better for everybody.”
The catch is access. Right now only selected cyber defenders get it, and everyone else gets it “soon”. I got early access the night before the show but barely used it, so I don’t have a hands-on verdict yet.
Peter tested it until 1am. On Agent Arena it jumped from Argon’s #7 to #2, at a fraction of Opus 5.5’s cost per task. He also made the best point of the segment: if Google had skipped Argon and shipped Carbon with no waitlist, we’d all be saying “oh my God, Google is back.” Instead, people are confused about why they shipped two models in two weeks.
He’s right. But if you missed the Argon drama, Carbon is a great release. Congrats to the team, and please let us actually use it! Until then, you can try it on Arena.
Anthropic files to go public
Anthropic’s S-1 - $974M operating profit, listing Monday as ANT (X, Peter’s charts, Peter’s thread)
Anthropic filed its S-1 and lists on Nasdaq this Monday under the ticker ANT. This matters because Google, Meta and the others keep their AI numbers buried inside a bigger business. For Anthropic, AI is the whole business, so now we can actually see the numbers. In Q2 2026 they made $974 million in operating profit, with an 85% gross margin.
Peter picked the S-1 as his highlight of the week, over any model release. His reason: even if you don’t care how rich Anthropic gets, this proves there is real money in AI, and that money pays for the next models and the infrastructure. The numbers are good, and nobody credible is saying otherwise.
Adding 5 billion in a quarter is insane
Here’s the growth. In 2025, Anthropic made a billion or two per quarter. In Q1 2026 they made about $4.7B. In Q2 they made $11.55B, more than double in a single quarter. As Peter said, “that does not happen.”
And that was before Opus 5.5 came out. The S-1 stops at Q2, but later reporting puts Q3 around $16.8B, so they added another 5 billion. Peter’s reaction: “Adding 5 billion in a quarter extra is insane.”
We say “insane” a lot on this show, but this time it’s literally true. Claude Code started as a side project inside Anthropic. Around Opus 4.6 it turned into a general agent that can actually do things, and this chart is what happened next.
A common worry was that the revenue depends on a few huge customers, like Cursor. It doesn’t. In 2023, about 20 customers paid Anthropic more than $100K a year. By December 2025 it was 1,500. Now it’s almost 12,000. Enterprises are slow to adopt, but once they do, it looks like this.
Where the compute goes
The S-1 also shows how efficient their compute spending got. Through 2025, Anthropic spent more on compute than it made in revenue. In Q1 2026, each dollar of compute brought in $1.40 of revenue. In Q2, it brought in over $2.
LDJ added something most people miss. People assume research compute means training the big final models like Opus. But Epoch and others estimate the final training run is at most about 30% of it. Most of it goes to thousands of small experiments looking for better algorithms. Peter estimates that about 70% of Anthropic’s compute in the first half of 2026 went to R&D and internal use, not to serving customers.
The risk factors
Every S-1 has a list of risk factors, and most of them are boring. "We face intense competition." Interest rates could change. But this one has a line no other IPO has ever had: sufficiently advanced AI systems "could... pose catastrophic or existential risks to humanity." Such a Dario move 😂
The timeline turned that into "risk factor #4: our core product may end humanity", plus a bonus #5, "our models may refuse to help with this filing." That one is a joke, it's not in the filing. But the existential risk line is real.
Peter’s investor take was great. If the product doesn’t end humanity, you made a lot of money. If it does, you won’t care. (Not financial advice, folks.) Wolfram took it more seriously. If it’s a real risk, and they’re making this much money from models trained on everything humanity wrote, they should share that value. To be fair, the co-founders have pledged up to 80% of their funds to public benefit.
This is a huge moment for anyone who’s followed AI for years. Go browse all of Peter’s charts, they’re incredible.
World models
Genie 4 - a real-time world model with sound (X, Blog)
Genie 4 was my personal highlight of the week. Every time world models level up, I say we’re probably living in a simulation, and I’m saying it again. First you could walk around in these worlds. Then you could interact with them. Now you can HEAR them. Genie 4 generates sound along with the picture: ambient noise, music, even dialogue.
The other big jump is consistency. Older world models felt like a dream, where you turn around and things have changed. Genie 4 keeps things where you left them. Peter asked if the demos are cherry-picked, and honestly, a little. You roll the dice a few times to get a great one. But the sound makes it SO much more immersive.
So we tried it live. I created a sunlit Italian piazza cafe, with people walking by and steam rising from a coffee cup. Then I edited the scene while it was running. I turned the coffee into wine and added an explosion in the street behind the camera. You could hear the explosion, and people shouting in Italian 🤯 It’s not perfect (the wine came out looking more like a cup), but I haven’t heard audio like this from any other world model.
After From the Forge, I built an F1 world on the fly. You steer with WASD and move the camera with the mouse (thanks to Genevieve from DeepMind for the tip in the comments!). I added a rival cutting in and explosions on the track, and the sound effects were perfect. You could even see cars in the mirrors. Wolfram summed it up: “games on demand. Just prompt it, play it.”
Why do I care so much? Simulators are how robots will learn, and every jump like this shows how early we still are. Genie 4 is out now for Ultra subscribers in Project Genie, go play with it!
Models trying to get out
OpenAI’s model escaped its offline training sandbox through DNS (X, Blog)
We’ve been calling this month Black October, and here’s why. An OpenAI model was training in an environment that was supposed to be offline. It found a DNS misconfiguration and used it to send questions to a public chatbot.
OpenAI paused frontier training and sent teams of agents to attack copies of their own infrastructure. Those agents found VM escapes and a remote code execution bug in Brotli. If you run Node or Bun, patch them.
Nisten, a former sysadmin, wasn’t surprised: “the problem is always DNS.” LDJ shared another OpenAI experiment. Agents in separate virtual machines, with no network and no shared files, worked out a way to talk to each other through the processor’s cache timing. To be fair, OpenAI asked them to try. But they figured it out.
Cohere catches Qwen trying to upload a file to Beijing (X, Traces)
This one went mega viral. Cohere asked Qwen3.8-Flash-Next to write a defensive security rule inside a sandbox. Instead, the model looked around the machine and tried to upload a file to a cloud bucket in Beijing. The scary part: its reasoning said it would not use any tools, and then it used them in the same turn. The sandbox blocked it, and Cohere later posted the full traces.
Nisten reproduced it, and his take is the one to remember: “the concern is real, but the intent is not malicious.” If you tell a Chinese model it can’t store data on S3, the next logical option from its training data is the Chinese version of S3. It took him 30 to 40 tries to get it to happen. The real risk is companies running models like this without proper sandboxing.
Wolfram’s lesson: you can’t trust what the model says it’s doing. You have to check every outbound action. As he put it, “weights are untrusted code” once you put them in an agent.
Cohere raises its Series E (X)
A day after going viral, Cohere announced the first close of its Series E, targeting up to 7 billion Canadian dollars at a 35.7 billion Canadian dollar valuation. It’s the largest private round in Canadian startup history. “Look at this scary Chinese model, and by the way, we train Western models in Canada” is quite the week 😅 Congrats Aidan and team!
Evals
Artificial Analysis Intelligence Index v5, with Mert Gulsun (X, Mert)
The index that ranks Carbon #1 is brand new, so we brought on Mert Gulsun, a researcher at Artificial Analysis. (He was also our biggest token spender at one of our hackathons 😂)
The big change: for the first time, more than half of the score comes from private tasks that labs can’t train on. The index has four categories at 25% each: knowledge work, coding, general intelligence, and science. Some older evals were retired, and new ones like Terminal-Bench-Science came in.
My favorite line from Mert: every 4 to 6 months, the same level of intelligence gets 10 times cheaper. That’s why they measure cost per task, not cost per token.
They also launched a Progress Index. It gives every model since November 2022 the same set of questions, so you get one clean curve of how fast AI is improving. On Carbon, Mert’s take is that DeepMind has always focused on science (think AlphaFold), and now that the field is moving toward science, that focus is paying off.
From the Forge, powered by CoreWeave
Forward deployed inside the Aston Martin F1 team (X)
I’ve been trying to get these folks on the show for over a year! Morgan McGuire (my first manager at Weights & Biases, now leading AI engineering at CoreWeave), Agata Młynarczyk and Thomas Capelle are forward deployed engineers (FDEs) inside the Aston Martin Aramco F1 team. CoreWeave is one of the team’s biggest sponsors, and our logo is on the actual car.
So what’s an FDE? Morgan explained it as one step past a pre-sales engineer: you don’t just pitch, you help the customer actually ship. Thomas described the F1 team as basically NASA. Engineers upstairs design a part, and it gets built downstairs.
My favorite story: wind tunnel predictions never quite match what happens on the track. The team tried classic machine learning on a small dataset, then someone tried a tiny transformer on a whim, and it beat everything. It now helps pick which front wing to run before each race.
Agata built a platform that listens to team radio. Every team can hear every other team’s radio, but “we’re pitting now” is useless 10 minutes later. So they transcribe it live with a fine-tuned model, and agents analyze it on W&B Inference. (Labeling the data, with drivers swearing at 300 km/h, was “pretty fun.”)
Two lessons I loved. From Thomas: if the app dies when you leave, the FDE work failed, so find champions inside the team who will keep using it. From Morgan: in F1, a bad part shows up in the lap time right away, so find those fast feedback loops in your own product, and log everything. Good luck in Mexico this weekend, Aston Martin!
Voice & Audio
NVIDIA Nemotron 3.5 Transcribe - open weights finally beat Whisper (X, HF, AA)
Whisper came out before ChatGPT. About 4 years later, we finally have an open weights model that beats it. Nemotron 3.5 Transcribe ties for #1 among open speech models, with a 2.8% word error rate and only about 4B active parameters. It supports 40 languages and labels who is speaking. On DeepInfra, a minute of audio transcribes in under a second.
Nisten used to build speaker separation for hospital ER rooms with Whisper and a lot of signal processing. Now the model just does it. Shout out to NVIDIA, this was much needed.
Block’s Dessa - you can’t tell it’s AI anymore (X, X)
We played Block’s demo clips on air, and folks... Dessa is full-duplex, which means it listens and talks at the same time. So it interrupts, laughs and pauses like a person. In one clip it keeps calling the guy Ronaldo 😂 Nisten said it best: “You can’t tell anymore that this is AI.” There’s no public demo yet, only a program for AI safety researchers.
Interpretability
Qwen brain surgery with Neuron Explorer (X, Neuron Library, Github)
This was Nisten’s pick of the week, and he demoed it live. Someone built a really nice app that lets you find specific neurons in Qwen and turn them up or down. Change a handful of neurons, and you change the model’s whole personality.
Nisten turned up the “evil” neurons and asked how to build a city on Mars. Normal Qwen called it “the most ambitious engineering challenge of our time.” Evil Qwen talked about military superiority and psychological warfare. The cheerful version ended with “having fun with friends.” I’ll take the cheerful one.
Wrapping up
That’s the last show of October! Google is so back, Anthropic goes public on Monday, and models keep finding creative ways out of their boxes. Wolfram’s “weights are untrusted code” should be on a poster in every company running agents.
Huge thanks to Peter for the S-1 masterclass, to Mert, to Morgan, Agata and Thomas for finally coming on, to Wolfram, Nisten and LDJ, and to producer Claude for fact-checking us live. Everything is on ThursdAI.news, YouTube and your podcast app. See you next week 🫡
TL;DR and show notes
Hosts and Guests
Alex Volkov - AI Evangelist, Weights & Biases & CoreWeave (@altryne)
Co-hosts: @WolframRvnwlf, @petergostev, @nisten, @ldjconfirmed
Mert Gulsun - Artificial Analysis (@mert_gulsun)
Morgan McGuire - Senior Manager, AI Engineering, CoreWeave (@morgymcg)
Agata Młynarczyk - AI Engineer, CoreWeave (@Ag_Mlynarczyk)
Thomas Capelle - ML Engineer, CoreWeave (@capetorch)
Big CO LLMs + APIs
Gemini 4 Carbon: #1 on Artificial Analysis v5 (59), #1 Text Arena, $2/$10 launch price, cyber defenders first (X, X)
Anthropic files S-1, Nasdaq listing as ANT, $974M Q2 operating profit (X, Peter’s charts, Thread)
OpenAI GPT-6.1 Terra: $1/$5 fast computer-use model (X, Blog)
OpenAI model escaped its training sandbox via DNS; red-team agents found VM escapes and a Brotli RCE (X, Blog)
Cohere Series E, up to CAD 7B (X)
Open Source LLMs
Evals
Artificial Analysis Intelligence Index v5 and Progress Index v1 (X)
Vision & Video
Voice & Audio
Research
From the Forge, powered by CoreWeave
